预训练模型在不同深度学习框架中的转换是一种常见的任务。今天刚好DPN预训练模型转换问题,顺手将这个过程记录一下。
核心转换函数如下所示:
def convert_from_mxnet(model, checkpoint_prefix, debug=False): _, mxnet_weights, mxnet_aux = mxnet.model.load_checkpoint(checkpoint_prefix, 0) remapped_state = {} for state_key in model.state_dict().keys(): k = state_key.split('.') aux = False mxnet_key = '' if k[0] == 'features': if k[1] == 'conv1_1': # input block mxnet_key += 'conv1_x_1__' if k[2] == 'bn': mxnet_key += 'relu-sp__bn_' aux, key_add = _convert_bn(k[3]) mxnet_key += key_add else: assert k[3] == 'weight' mxnet_key += 'conv_' + k[3] elif k[1] == 'conv5_bn_ac': # bn + ac at end of features block mxnet_key += 'conv5_x_x__relu-sp__bn_' assert k[2] == 'bn' aux, key_add = _convert_bn(k[3]) mxnet_key += key_add else: # middle blocks if model.b and 'c1x1_c' in k[2]: bc_block = True # b-variant split c-block special treatment else: bc_block = False ck = k[1].split('_') mxnet_key += ck[0] + '_x__' + ck[1] + '_' ck = k[2].split('_') mxnet_key += ck[0] + '-' + ck[1] if ck[1] == 'w' and len(ck) > 2: mxnet_key += '(s/2)' if ck[2] == 's2' else '(s/1)' mxnet_key += '__' if k[3] == 'bn': mxnet_key += 'bn_' if bc_block else 'bn__bn_' aux, key_add = _convert_bn(k[4]) mxnet_key += key_add else: ki = 3 if bc_block else 4 assert k[ki] == 'weight' mxnet_key += 'conv_' + k[ki] elif k[0] == 'classifier': if 'fc6-1k_weight' in mxnet_weights: mxnet_key += 'fc6-1k_' else: mxnet_key += 'fc6_' mxnet_key += k[1] else: assert False, 'Unexpected token' if debug: print(mxnet_key, '=> ', state_key, end=' ') mxnet_array = mxnet_aux[mxnet_key] if aux else mxnet_weights[mxnet_key] torch_tensor = torch.from_numpy(mxnet_array.asnumpy()) if k[0] == 'classifier' and k[1] == 'weight': torch_tensor = torch_tensor.view(torch_tensor.size() + (1, 1)) remapped_state[state_key] = torch_tensor if debug: print(list(torch_tensor.size()), torch_tensor.mean(), torch_tensor.std()) model.load_state_dict(remapped_state) return model
从中可以看出,其转换步骤如下:
(1)创建pytorch的网络结构模型,设为model
(2)利用mxnet来读取其存储的预训练模型,得到mxnet_weights;
(3)遍历加载后模型mxnet_weights的state_dict().keys
(4)对一些指定的key值,需要进行相应的处理和转换
(5)对修改键名之后的key利用numpy之间的转换来实现加载。
为了实现上述转换,首先pip安装mxnet,现在新版的mxnet安装还是非常方便的。
第二步,运行转换程序,实现预训练模型的转换。
可以看到在相当的文件夹下已经出现了转换后的模型。